You: "Deploy this service to GKE"

─── Step 1: the agent searches ───────────────────────────────────────

Agent: search_skills("deploy a containerised service on kubernetes")   ← 6ms
     → 5 results (summaries only, no full instructions):
       1. "gke-service-networking"   (0.56) - Gateway API, Ingress, Cloud Armor, NEGs, managed SSL
       2. "gke-workload-scaling"     (0.51) - HPA and VPA for GKE workloads
       3. "gke-manifest-generation"  (0.51) - Production-ready Kubernetes YAML for Autopilot/Standard
       4. "gke-app-onboarding"       (0.46) - Containerizing and deploying an app to GKE for the first time
       5. "gke-basics"               (0.44) - Cluster provisioning, credentials, Autopilot vs Standard

─── Step 2: it reads the descriptions and picks #4, not #1 ───────────

Agent: get_skill("gke-app-onboarding")
     → gets the full guide: containerization, manifests, migration path
     → writes the Dockerfile and deployment.yaml

─── Step 3: a new need emerges mid-task ──────────────────────────────

Agent: # the service has to survive traffic spikes — search again
       search_skills("autoscale pods on cpu and memory")               ← 6ms
     → "gke-workload-scaling" (0.61) - Horizontal and Vertical Pod Autoscaler for GKE
     → reads the guide, adds the HPA manifest

pip install "skill-retrieval-mcp[local,hf]"
skill-mcp pull --include-index      # corpus + pre-built vector index
skill-mcp init                      # detect and register with your agents

{
  "mcpServers": {
    "skill-retrieval": {
      "command": "/absolute/path/to/skill-mcp",
      "args": ["--data-dir", "/absolute/path/to/data-dir", "serve"]
    }
  }
}

<!-- ~/my-skills/deploy-checklist/SKILL.md -->
---
name: "deploy-checklist"
description: "Pre-deployment verification checklist for production releases"
tags: ["deployment", "production", "checklist"]
---

## Steps

1. Run full test suite...
2. Check database migrations...

skill-mcp import --source directory --path ~/my-skills/

~/.skill-mcp/
├── config.yaml
├── skills.db          # SQLite + FTS5
└── index/             # FAISS

skill-mcp --data-dir /srv/skills pull

# set backend: openai, model: text-embedding-3-large in config.yaml, then
skill-mcp build-index --backend openai

skill-mcp [--data-dir DIR] [--log-level LEVEL] COMMAND [ARGS]

  init [--data-dir DIR] [--no-register]        Set up the data directory, register with agents
  pull [--replace] [--include-index]           Download the corpus from HuggingFace
  import --source SOURCE --path PATH           Import your own skills
       [--no-index]
  build-index [--backend B] [--model M]        Build or update the vector index
       [--force]
  serve [--transport stdio|sse]                Start the MCP server
  search QUERY [--k N]                         Search from the terminal
  status                                       Show what is loaded
  dedup                                        Remove cross-source duplicates

git clone https://github.com/JayCheng113/skill-retrieval-mcp
cd skill-retrieval-mcp
pip install -e ".[all,dev]"
pytest tests/ -v    # 240 tests, ~6s